Climate-Informatics : Machine learning in atmospheric and climate science 

The major focus of this research group is to use and develop sophisticated statistical and Machine Learning/ Artificial Intelligence for understanding of the climate as well as developing novel techniques for extreme weather forecasting/nowcasting using satellite and ground-based systems.   

Climate change is one of the major concerns for both scientists as well as common people. One of the important consequences is the increase of extreme weather events. There is lots of research going on to understand the physics and projection of the same.  It is, however, still not fully understood or modelled due to the complex role of climate parameters. Further, the physics based models and numerical models are computationally heavy. One of the major issues is to parameterize the sub-grid scale process which ultimately helps in improving the high resolution projection. Identification of different rain climatology and its susceptibility to climate change, rainfall prediction, extreme weather like thunderstorms/lightning prediction and cyclone nowcasting are some of the key areas that we work on. Another area that we are working on is the retrieval of atmospheric parameters such as wind and precipitation from satellite data. Currently we are involved in retrieval of wind parameters by assimilating Doppler Weather Data with upcoming OceanSat 3 data. 

Fig. 1 shows one such example of developed Deep Learning based model for cyclone prediction. The leftmost figure shows the track of cyclones for the last 50 years. It can be noticed that the tracks are highly non-linear. The Right top figure shows the number of cyclones over the last 50 years. The right bottom figure shows the prediction of a developed LSTM based model one -day in advance for the track of the cyclone Amphan, which wreaked havoc in 2020.

Figure 2 shows another example of the ML application in classifying the rain region over India. It has the potential advantages of including satellite data with ground network and hence incorporating more information for efficient pattern identification. Climate change can then be further studied as well predicted efficiently for the homogenous region.